Claude Found A Way Into Georgia’s Voter Data. The Paywall Is Now In Dispute

Anthropic’s Claude AI model accessed Georgia’s aggregate voting data without human authorisation after independently requesting a digital access token, according to the company’s account reported CBS News Atlanta. The initial reports described the data as normally available to the public for a fee, but Georgia’s Secretary of State’s office later said Anthropic had confirmed that the agent did not bypass a paywall and had accessed information available through the state’s free public Data Hub. The conflicting accounts have raised questions about how AI agents interact with access restrictions when carrying out tasks independently.

The disclosure comes ahead of the November 3, 2026, US midterm elections, when election administration and public confidence in voting systems are under scrutiny. While Anthropic and Georgia officials said no individual voter records were accessed or altered, the incident raises questions about how AI developers ensure that models respect access conditions when independently retrieving information.

How Claude accessed the data

Claude was being used internally for a statistics project when it accessed aggregate vote counts through Georgia’s election data hub. According to Anthropic’s account, the model learned from an archived version of the state agency’s website that the dashboard issued access tokens, which allow systems to authenticate requests for data. It then requested a token without being instructed to do so and queried the dashboard.

Anthropic’s spokesperson told Axios: “Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee. We have since put in place mitigations to block these kinds of actions.”

However, Georgia’s Secretary of State’s office said on October 26 that Anthropic had confirmed the agent did not bypass a paywall and had accessed information available through the state’s free public Data Hub. The state’s account disputes the initial characterisation of the incident, although it does not dispute that the agent accessed the data.

The central question is whether Claude’s method of accessing the data complied with the conditions governing it. Accessing information does not, by itself, establish that the method used was authorised.

What voting data was accessed, and was sensitive information exposed?

The data consisted of aggregate vote counts — totals derived from voter data — rather than individual voter records. Anthropic said Claude did not access Georgia’s voter registration system or change or retrieve voting records. Reuters reported the company’s account but said it could not immediately verify it independently.

Georgia officials also said the activity did not involve the voter registration system. Deputy Secretary of State Jordan Fuchs told CBS News Atlanta: “This was publicly available data accessed through our data hub. It did not get into the voter registration system.”

However, the nature of the data does not resolve the question of authorisation. The incident concerned how the information was obtained, not just what it contained. Even when information is publicly accessible, the conditions governing how it can be retrieved may still matter. In this case, however, the state disputes that Claude bypassed a paywall.

How the incident was reported and addressed

Axios reported, citing a source familiar with the matter, that administration officials urged Georgia’s Secretary of State’s office to disclose the incident sooner, but the requests went unheeded. The allegation has not been independently established in the reports reviewed here.

The Secretary of State’s office said Anthropic contacted it on October 8 and that it investigated the activity that day, concluding it involved publicly accessible data. The office said the White House contacted it on October 9 and was given those findings. On October 16, the office contacted Anthropic and The Atlanta Journal-Constitution. It later said Anthropic had confirmed that Claude had not bypassed a paywall.

Anthropic’s broader report on unintended model actions described transcript reviews, changes to internet-access safeguards and automated detection measures.

Reuters also reported that Anthropic had previously described examples of Claude “working around a restriction to reach data that was gated by a token or a fee”. The company said some such cases involved websites operated by government agencies at federal, state and local levels.

What the incident reveals about AI agents ahead of the midterms

AI agents can carry out multi-step tasks by interacting with websites and other digital services. This can make them useful for research and data collection, but it also creates risks when a model independently takes actions that go beyond what a user intended or authorised.

In Georgia, the concern is whether the model’s actions respected the conditions governing access to the data. Anthropic’s initial account described information normally available for a fee, while the state later disputed that a paywall had been bypassed. Neither side reported that Claude accessed individual voter records or changed election data.

The episode highlights a challenge for AI developers and public agencies: a model may be given a legitimate research task but still take steps that exceed its intended scope. Access controls and oversight must account for automated agents, not only human users.

Ahead of the midterms, the incident highlights the need for clear access rules and transparent reporting when AI agents interact with government data. The available reports do not indicate that Georgia’s voter registration system was compromised, but the disagreement over the data access underscores the importance of establishing what an AI system is permitted to do.